LLMs Struggle to Replicate Cross-Cultural Depression Symptom Patterns

Clinical psychology research has long documented a cultural divide in how depression manifests: Western patients tend to emphasize psychological symptoms like mood disturbances and cognitive issues, while Eastern patients more often report somatic (physical) symptoms like pain and fatigue. A new paper on arXiv investigates whether large language models—which are increasingly deployed in mental health contexts—reproduce these culturally grounded patterns.

The Study#

Researchers tested LLMs by prompting them with Western or Eastern personas to see whether the models would generate symptom descriptions aligned with clinical norms. The findings were sobering: LLMs largely failed to replicate established cultural patterns when prompted in English. However, prompting in major Eastern languages (Chinese, Japanese, and Hindi) showed modest improvements in capturing the expected regional symptom expressions.

Why Models Fall Short#

The analysis identified two core problems:

• Low sensitivity to cultural personas: Models don't meaningfully adjust their outputs based on cultural identity signals in prompts.
• Culturally invariant symptom hierarchy: The models appear to have a fixed, deeply ingrained ranking of symptoms that overrides cultural cues, suggesting they've absorbed a dominant (likely Western-centric) framing from training data.

Implications for Mental Health#

The research underscores a critical gap: while prompt language matters, general-purpose LLMs lack the robust, culture-aware capabilities needed for safe and effective mental health applications. As these models are increasingly used for screening, diagnosis support, and patient communication, the inability to recognize culturally specific symptom expressions could lead to misdiagnosis or inadequate care—particularly for non-Western populations.